Desktop app for Claude Code and Codex — Caprock

Free desktop app for coding agentsRun your coding agents. Follow every PR to release.

Runs entirely on your machine. No account, no cloud, no telemetry.

  • Claude Code
  • OpenCode
  • Codex
  • Gemini CLI
  • DeepSeek Harness
Download for Mac

Notarized by Apple. One file for Apple silicon and Intel, macOS 13 or later.

Or run it yourself:

curl -fsSL https://caprock.dev/mac.sh | sh

Paste in Terminal: it installs Caprock and opens it.Or with Homebrew: brew install --cask dspv/tap/caprock-app

All install options →

In August I ran $5,898 of Claude Code on a $200 Max plan — 29× at API list prices.

Caprock on a phone: today's cost at API list price, sessions and turns, and the Claude Max 5-hour and weekly windowsCaprock on a phone: today's cost at API list price, sessions and turns, and the Claude Max 5-hour and weekly windows

The macOS app replaying a real Caprock work session, and a paired phone. Example figures.

What did that feature cost?

Cost per pull request, with the tokens behind it.

The pull requests open now or merged this week that cost the most, each priced from the agent sessions on its branch. Not one total for the month.

At API list price: an estimate of what those tokens cost, not a bill. Example figures from a demo machine.

A real week, read on the maker's machine 16 Oct 2026: 45 pull requests merged into dspv/caprock in 7 days, $343.78 of agent cost — about $7.64 each; the dearest, #720, $71.13. At API list price.

On the Home screen, with GitHub connected.

Pipeline

Every pull request, from agent to release.

Read from your git and GitHub, on your machine: where each pull request is, what it cost, and what went out in each version.

Open, CI, merged, released, and what the agents spent on it.

Example figures from a demo machine. Agent cost is at API list price, an estimate.

Work

Who is on what, and how far it got.

Every branch at its furthest stage, with the agent and model on it, how long it has run and what it has cost so far.

Example figures from a demo machine. Agent cost is at API list price, an estimate.

Your agents

Your agents and shells, in one window.

Claude Code, Codex, OpenCode, Gemini CLI and Hermes Agent in tabs and split panes, each in its project or worktree.

Sessions waiting on you come first; a new agent is one line away.

Claude Code desktop app options →

Example figures.

Cost and limits

What each repository costs, and how close the limit is.

Every turn priced at API list price, live, with your Claude and Codex plan windows beside it.

Cost

Claude Code cost per repository — and per service in a monorepo.

A turn counts toward the directory of the last file it touched, so the rows add up to the repository.

127.0.0.1:22776
Projects30d
example-app
41.2M
$3,426
by files touched · share of this repository’s tokens
/app
25.1M 61% · $2,090
/.ai
2.8M 6.8% · $233
/app/tests
1.1M 2.8% · $96
outside the repository
10.7M 26% · $891

A turn counts toward the directory of the most recent file it touched, and keeps counting there until work moves elsewhere. Each turn goes whole to one row, so the rows add up to the repository’s total exactly.

Example figures.

See what your agents did today.

Free and open source · macOS, Windows and Linux · no account

View on GitHub →
Download for Mac

Notarized by Apple. One file for Apple silicon and Intel, macOS 13 or later.

Or run it yourself:

curl -fsSL https://caprock.dev/mac.sh | sh

Paste in Terminal: it installs Caprock and opens it.Or with Homebrew: brew install --cask dspv/tap/caprock-app

Private by design

Your sessions stay on your machine.

No account, no telemetry. Network features, a paired phone among them, are each off until you turn them on.

Your computer

127.0.0.1:22776caprock.db · SQLite

Off until you turn it on

Release check · offGitHub · offPhone · offTelegram · off

Local by default

Data stays in a SQLite file on your disk. There is no account or telemetry. It goes online only for what you switch on: update checks, GitHub, a paired phone, your own Telegram bot.

~30 MB

The daemon, idle.

~40 MB while reading 1,414 transcripts. Measured 2026-08-18.

hook fails → exits in < 1 s

Hooks exit safely on error

If the hook fails, it exits silently within a second so your session can continue.

your terminal · observedstarted by Caprock · controlled

You control your terminal sessions

Caprock only observes sessions you start yourself. It controls only the ones it starts.

priced from a dated list

No invented numbers

Prices are sourced with a date; a forecast is labelled a forecast.

macOSWindowsLinux

Three operating systems

macOS, Linux and Windows, tested on all three from day one.

Compared

Caprock and other apps for running coding agents.

Each runs agents in parallel. What differs is the price and licence, the platforms, how a phone reaches your machine, and what you see of cost and limits.

✓ yes · ◐ partly · — no, or not in their docs. Caprock, Orca, Conductor, Warp and Nimbalyst run both Claude Code and Codex; the Claude app runs Claude Code only, the ChatGPT app Codex only. From each maker's own docs, pricing pages and repositories, read 7 October 2026; cells that changed since, re-read 21 October 2026. All columns, with sources → · Claude Code GUIs →

Claude Code compared with Codex · Cursor · OpenCode · Gemini CLI · Copilot · Antigravity · Kiro

  • Caprock fits

    You run several agents across several repositories and want their cost, plan limits and history in one place, with a phone that reaches them from anywhere and no Caprock account.
  • Look elsewhere

    You want a code editor or cloud workspaces: Caprock runs your own agents in your own terminals, on your machine.

Try it

On the sessions you already run.

Install the app and it picks up the sessions already on this machine, past ones too.

Notarized by Apple, for Apple silicon and Intel. Other ways →

Caprock with six Claude Code agents in tabs, a shell under the running one, and its cost on the rightCaprock with six Claude Code agents in tabs, a shell under the running one, and its cost on the right

Your record

Keep your coding agents’ history.

Caprock imports the sessions still on disk, not ones already deleted, and records new ones. You can search them after your agents delete the original transcripts.

127.0.0.1:22776
Last 30 days
$342.09
at API list prices · 3 models
opus-4.8 · $270
sonnet-4.6 · $50
haiku-4.5 · $22

Example figures.

  • Kept past 30 days

    Claude Code deletes transcripts after 30 days by default. On the maker's machine, 67 of 114 transcripts were gone, including 10 from worktrees he deleted. All 67 remained searchable in Caprock. The numbers →
  • Six agents, one table

    Claude Code, Codex, OpenCode, Gemini CLI, DeepSeek Harness and Hermes Agent each write their own format. Caprock reads all six into one table on your disk; caprock export gives it to you as TSV, CSV or JSON lines. How →
“not just tracking spend, but seeing what's really happening on your projects. … PyCharm restarts and the session is gone. Caprock takes that pain away.”
Vladimir Konov
Vladimir KonovSoftware Engineer · Rust, distributed systemsread the whole note
“I see its future value as a central, independent store of my sessions across different models, providers and projects … use as parallel documentation or extra memory.”
Sergey Aksёnov

Every note in full, with how long each person used it →

FAQ

Questions.

What does Caprock do?+

Caprock is a free desktop app for macOS, Windows and Linux to run and watch several coding agents — Claude Code, Codex, OpenCode, Gemini CLI and others — in one window, with what each session costs. Each session gets a terminal tab, with split panes, your projects and their worktrees in the sidebar, and the sessions waiting on you first. The dashboard is inside it: what every session on the machine is doing — including the claude you start in your own terminal — what it has cost per repository, your plan limits, and when one is stuck. Your phone opens the same Caprock once paired.

Is there a desktop app and a phone app?+

A desktop app, yes; the phone needs no app of its own. The desktop app runs on macOS (download it, or brew install --cask dspv/tap/caprock-app; signed with an Apple Developer ID and notarized by Apple), Windows and Linux, with the daemon — the background service that records your sessions — inside. It opens on your agent terminals and projects, and when a Claude Code session Caprock started asks for permission, the notification on macOS and Linux has Deny, and Approve when it shows the whole request. A phone pairs by QR code and opens Caprock in its browser, from anywhere, through an end-to-end encrypted relay: it reads a session as a chat and reconnects by itself, and once you allow control it answers prompts, clones a project and starts an agent in it.

Do I have to change how I work?+

No. Keep running claude in your own terminal. Caprock adds a small hook to ~/.claude/settings.json — backed up first, removed with caprock hooks uninstall — and reads the transcripts Claude Code already writes.

Will it see sessions from before I installed it?+

Yes. Caprock imports transcripts still on disk, which Claude Code keeps for 30 days by default. It then records new sessions and keeps them after Claude Code deletes the original files.

Does anything leave my machine?+

Only if you enable a network feature. By default, data stays in a SQLite file on your disk and the dashboard listens on 127.0.0.1. There is no account or telemetry. You can enable update checks (the app asks once), connect GitHub to clone and open pull requests, send weekly reports and phone alerts to your Telegram bot, or ask Gemini using your API key. Updating downloads the new app from GitHub's release files. You can also pair a phone, which reaches Caprock from anywhere, through an end-to-end encrypted relay that sees only ciphertext. A phone you pair can control sessions unless you untick that when you pair it, and you can take control away in settings. A bug report or a survey answer reaches us only when you press Send. The code is Apache-2.0 on GitHub.

How does it spot a stuck session?+

When a session calls the same tool with nearly the same input five times within three minutes, a banner goes up. Repeated reads and searches don't count: re-reading a file is normal work, and on 64,733 real tool calls they were over half of all alerts. Sessions Caprock started can be paused automatically.

Is it free?+

Yes, for solo use, and what is free stays free: live sessions, cost per repository, plan-limit warnings, history and search. A licence — $30 a year or $100 once — adds 8 features: Auto-pause at a daily budget, Pause at the plan limit, Answer from Telegram, A weekly report, Ask about your usage, Where each tool failed, Custom agents, More than three sign-ins per agent. The desktop app is free too. Caprock adds nothing to your Claude bill. The costs it shows are at API list prices: on a flat plan, an equivalent, not a charge.

How is it different from ccusage?+

ccusage prints usage tables in your terminal when you ask, with nothing running in between. Caprock runs in the background: live sessions, loop alerts, cost per repository, and history kept past Claude Code's 30-day cleanup. Use ccusage for a quick table, Caprock to watch everything in one place.

What is the task runner?+

An opt-in way to run tasks unattended. You write a task and the commands that must pass, such as tests, type checks and lint. Each task runs in its own git worktree; Caprock runs your commands itself before marking the task done. It never merges into your branches. Workers skip permission prompts and your commands run with your permissions, unsandboxed — give it independent tasks you would run yourself.

For teams

See your team’s agent costs by repository and service.

Track costs by type of work and see live sessions across the team. The dashboard runs in your network. Laptops send usage figures; prompts and code stay on each laptop.

caprockfor teams7 people · 30d
By personby cost
  • anna$1,240
  • dmitry$970
  • sam$670
  • priya$390
  • jonas$220
By repositoryby cost
  • payments-core$1,510
  • web-client$940
  • data-pipeline$610
  • admin-panel$260
Running now4 sessions
  • anna running go test ./…
  • sam editing checkout.ts
  • dmitry same call ×14
  • priya waiting on you

Loop alerts show which session is repeating a tool call.

Example figures. Your team's own numbers replace them from the first day.

From $500 a month, billed yearly — one flat price for the team, never a share of your AI spend. The Community Edition stays free.

See what your agents are doing.

Free and local. Your coding agents in one window.

Download for Mac

Notarized by Apple. One file for Apple silicon and Intel, macOS 13 or later.

Or run it yourself:

curl -fsSL https://caprock.dev/mac.sh | sh

Paste in Terminal: it installs Caprock and opens it.Or with Homebrew: brew install --cask dspv/tap/caprock-app

All of that is free and open source. Premium pauses the agents it starts at a daily budget or before the plan limit, and lets you answer them from Telegram.